The Walk-Back Year
Through 2025, every frontier AI lab was selling AGI by 2027. In 2026 the same labs are quietly retiring the capability claim while collectively committing $725 billion to data-center construction. The walk-back is real, the capex is real, and the math that connects them does not work yet.
AGI Is "Not a Super Useful Term" Now
The most reliable way to track an industry's confidence in its own narrative is to watch what its CEOs stop saying. Through most of 2024 and into 2025, the senior leadership of OpenAI, Anthropic, Google DeepMind, and the smaller frontier labs spoke openly about Artificial General Intelligence — what it was, when it would arrive, and what the appropriate institutional response should be. The dates given ranged from 2026 to 2030. The confidence intervals were absurdly narrow for a scientific prediction. The funding rounds that followed were even narrower.
Through the first half of 2026, those same voices have changed register.
A Curated Reading of the Walk-Back
Sam Altman recently described AGI as "not a super useful term" because "everyone defines it differently." That is a different sentence from the one he was giving interviews in 2024.
Dario Amodei remains the most bullish of the major lab CEOs, still publicly pointing at 2026-2027 as the window for transformative capability — but his framing has shifted to a "country of geniuses in a data center" formulation that explicitly conditions the arrival on "breakthroughs that have not yet materialized."
Andrej Karpathy, a former OpenAI co-founder, recently said agents "aren't anywhere close" and put human-level AI a full decade out. Karpathy is not selling AGI subscriptions, but he was instrumental in defining what the field thought AGI meant in 2023.
The industry-wide replacement vocabulary is "agentic AI." Where 2024's pitch deck slide promised "AGI: a system that can perform any economically valuable task a human can perform," the 2026 equivalent promises a "digital colleague that completes specific workflows alongside human teammates." Those are profoundly different claims. The first promises a phase transition. The second promises a tool. Tools are useful. Tools are also valued differently by capital markets.
$725 Billion, Up 77 Percent
If the capability narrative is softening, the spending narrative is doing the opposite. The four U.S. hyperscalers — Microsoft, Alphabet, Meta, and Amazon — have collectively guided to $725 billion in 2026 capital expenditure, up from approximately $410 billion in 2025. That is a 77 percent year-over-year increase, in absolute dollar terms larger than the entire annual GDP of Switzerland, and almost all of it is allocated to data-center construction and the AI accelerators that go inside them.
Meta's full-year projection sits at $145 billion and rising. Microsoft's CFO attributed $25 billion of the Redmond figure to rising memory chip and component costs alone — meaning a substantial portion of the headline increase is not "more capacity" but "the same capacity at higher prices," a detail that ought to concern anyone who believes the AI infrastructure curve is bending toward cheaper compute.
The most arresting framing comes from the capex-to-revenue ratio. For these four companies — which generate approximately $1.6 trillion in annual revenue between them — capex is now running at 45 to 57 percent of revenue. There is no industry in modern economic history that has sustained that ratio for a sustained period outside of national-scale infrastructure build-outs and oil-and-gas megaprojects. Telecom at the peak of the 1999-2001 fiber buildout briefly touched 40 percent. The current AI capex regime is materially heavier.
The Two Stories That Have to Reconcile
The bull thesis says revenue catches up — Microsoft's AI business is at a $37 billion annual run rate, growing 123 percent year-over-year; Google Cloud is at $20 billion, growing 63 percent. The bear thesis says the capability ceiling means revenue plateaus at "AI as a useful tool" rather than "AI as a phase transition" — which is enough to amortize maybe a third of the announced capex. Both stories are operating on the same evidence right now. One of them is about to be very wrong.
What the Research Community Has Quietly Conceded
The intellectual edifice that justified the 2024-2025 capex trajectory was the scaling-laws literature — a series of empirical papers, most originating at OpenAI and Anthropic, showing that model loss decreased predictably with increases in parameters, training compute, and training data. Pour in more of each, get a smarter model. The mathematics looked clean enough to underwrite multi-hundred-billion-dollar capital allocation decisions, and it did.
Through late 2025 and early 2026, the same researchers started publishing papers acknowledging — usually obliquely, sometimes directly — that the empirical curves are bending. Making models larger no longer delivers proportional capability gains on benchmarks. High-quality training data sources are running out. The next-frontier architecture papers coming out of the major labs are noticeably less about scale and noticeably more about inference-time compute, retrieval-augmented context, tool use, and agentic orchestration.
Each of those four research vectors is real and useful. None of them is the same as the scaling-laws thesis. They are engineering improvements — clever ways to get more out of a given model — rather than scientific demonstrations that a fundamentally smarter model can be built by spending more on training.
The Architecture Tells You the Bet
The clearest single signal that the industry has internalized the walk-back is the architecture of GPT-5. Disregard the marketing copy. Look at the engineering decisions.
GPT-5 is not a single monolithic model. It is a router plus an ensemble: a small, fast "default" model that handles most queries, paired with a larger, more expensive "thinking" model that the router invokes for harder problems. The system selects which model to use on a per-query basis. From the user's perspective it looks like one model. From the cost-of-goods-sold perspective it is a substantial efficiency play.
What the Architecture Implies
When the headline next-generation product from the leading frontier lab is sold on the basis of faster, cheaper, more efficient routing of existing model capabilities rather than fundamentally smarter, the company has already conceded — internally — that the scaling-laws trajectory is no longer the route to the next leap. You do not engineer a routing layer between a small model and a large model unless your large model is too expensive to use as the default. That is a different problem from "the large model is amazingly capable." It is a deployment-economics problem, not a capability problem.
The GPT-5 release was met with measurable user disappointment relative to expectations. The capability uplift over GPT-4-class models was real but incremental — closer in flavor to a software-version-bump than a generational leap. That, combined with the routing architecture, tells the broader story. The industry's flagship product is now an efficient ensemble of models that aren't individually much smarter than their predecessors.
What Has to Be True for $725 Billion to Pencil Out
The strongest case for the current capex regime is a simple optionality argument. The frontier labs and the hyperscalers do not have to believe AGI arrives in 2027 to justify the build-out. They only have to believe that whoever owns the most compute at the moment any meaningful architectural advance arrives wins the future — and that the cost of being short on capacity dominates the cost of being long on capacity.
That argument is not crazy. It is also not what is being told to capital markets.
The argument being told to capital markets is that AI revenue is going to scale into the capex on a roughly four-to-five-year amortization schedule. Run the math on a 30 percent gross margin on cloud-AI compute and a four-year asset life on the GPUs, and that requires the hyperscalers, in aggregate, to be doing approximately $500 billion in annual AI revenue by 2029-2030. The 2026 aggregate AI revenue across the four companies is probably $80 to $100 billion — Microsoft's $37B, Google Cloud's pro-rata AI portion, AWS's Bedrock plus Trainium, and Meta's substantial-but-internally-consumed AI compute that doesn't show up as revenue but does show up as ad-targeting margin.
That requires a sustained 30 to 40 percent compound annual growth rate for four years. In the absence of a capability step-change — the one that's currently being walked back — that growth has to come from selling more of the existing capability to more enterprises at lower prices. The competitive dynamics of that scenario, with four hyperscalers and several frontier labs all running the same playbook, do not favor margin maintenance. Compute is becoming a commodity at exactly the moment the commodity supplier has to charge premium margins to justify the build.
This is the structural problem the walk-back creates. The capex was sized for the capability-as-phase-transition story. The revenue ramp has to come from the capability-as-tool story. The denominators don't match.
The Real Output of the $725 Billion
Stripping out the capability narrative entirely, what is the AI capex actually purchasing? Three answers, each worth examining on its merits.
1. Power-Constrained Compute
The binding constraint on data-center construction in 2026 is not GPUs. It is electricity transmission and substation interconnect. The hyperscalers have effectively bought up multi-year power-purchase agreements with every commercial-scale generator in markets where AI workloads can plausibly cluster. The capex is partly buying the GPUs and partly buying first-claim on grid capacity in Virginia, Texas, Ohio, and Arizona. The grid-capacity option is real and likely valuable regardless of AI capability trajectory.
2. Inference Capacity for "Agentic" Workloads
Agentic AI systems — the new replacement vocabulary for AGI — require dramatically more inference compute per task than chat-style interactions. A single agentic workflow can invoke a model dozens or hundreds of times before producing a result. If agentic workflows become the dominant deployment pattern (a real possibility, even without a capability step-change), the inference compute required to serve them grows by 1-2 orders of magnitude. The capex is partially a bet that this transition happens fast enough to absorb the capacity.
3. The Next-Architecture Option
If a fundamentally different model architecture emerges — world models, continual learning, neurosymbolic hybrids, or something not yet named — the hyperscaler with the most existing infrastructure has a substantial first-mover advantage in operationalizing it. This is the optionality argument, and it requires no particular timeline on the architecture appearing. The capex is partly a perpetual option on "the next thing," priced at whatever it costs to keep building.
Each of those three justifications has merit. The composite is plausible. What's no longer plausible is the original story: that the existing transformer-plus-scaling trajectory delivers AGI within the asset lives of the data centers currently under construction. That story has been retired by the people telling it.
What to Watch in the Next Four Quarters
The Walk-Back Without the Capex Cut
It is unusual for an industry to retire its central marketing claim — AGI — at the same moment it is committing to the largest infrastructure build in private-sector history. Both things are happening simultaneously, in 2026, at the same companies. They are not being reconciled out loud.
The reconciliation is probably one of three things. The first is that the labs are wrong about the slowdown and the next architectural advance is closer than the public posture suggests — in which case the capex looks prescient in retrospect. The second is that the optionality argument is the real argument and the AGI marketing was always a fundraising story — in which case the capex makes sense even though the public narrative has to shift to "tool, not phase transition." The third is that the capex is wrong, the revenue ramp doesn't happen, and 2027-2028 brings a hyperscaler write-down cycle comparable in scale to the 2001-2002 telecom equipment write-offs.
All three outcomes have advocates inside the same companies, often in the same conference rooms. The honest answer is that nobody knows which one is right, including the people committing $725 billion against it. The walk-back is real. The capex is real. The mismatch is the actual story.
The most expensive thing you can do during a capability plateau is build more capacity. The second-most expensive thing you can do is build less.
- Big Tech 2026 capex reaches $725 billion, up 77% from last year — Tom's Hardware. Source of the $725B headline figure and the company-level breakdowns.
- Skyrocketing component prices push Big Tech capex to record $725 billion — Microsoft alone attributes $25 billion to memory and chip costs — Tom's Hardware. The component-pricing footnote on the capex line.
- Hyperscalers Hit $700 Billion in 2026 AI Spending Plans — Yahoo Finance.
- Big Tech Q1 2026 Earnings Power $700B AI Capex Spree — HeyGoTrade. Capex-to-revenue ratios (45-57%).
- Tech AI spending approaches $700 billion in 2026, cash taking big hit — CNBC, February 2026.
- Big Tech's $700 billion AI spending spree has no clear end in sight — Fortune, April 30, 2026.
- AI Beyond the Scaling Laws — HEC Paris. Academic treatment of the diminishing-returns inflection.
- AGI Still Years Away, Despite Tech Leaders' Bold Promises for 2026 — Cogni Down Under, Medium. Source of the Karpathy quotation and the Altman "not a super useful term" framing.
- State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI — Lex Fridman Podcast #490, transcript. Multi-hour discussion among researchers on the current capability ceiling.
- Beyond the Scaling Laws: Why the Next Leap in AI Requires an Architectural Revolution — Algorithma.
- AGI's Last Bottlenecks — AI Frontiers.
- What I Learned About Hyperscalers' AI Spend — Om Malik, April 30, 2026. Independent journalist's read on the Q1 earnings cycle.
About Hard-Ceilings
Hard-Ceilings is Equicurious's technology-assessment desk. We cover the gap between what a technology is sold as and what it can actually do. Our standing assumption is that the cost of being wrong about the second question, when capital allocation depends on the first, is paid by people who don't get to be in the room when the assumptions get retired.
Equicurious provides educational content only, not investment advice. The framing above is interpretive; the underlying capex figures and CEO quotations are from public filings, earnings calls, and on-the-record interviews cited in the sources panel. Reasonable observers reach materially different conclusions about whether the current capex regime is prescient or excessive. Past performance does not guarantee future results.